Designers in the repo
Jack Anglesea · 5 October 2026
Read the originalJack Anglesea of Mettle Studio describes building ten workshop prototypes in code with an AI agent, and recording each expert decision in a file beside the prototype.
Jack Anglesea is a lead designer at Mettle Studio. His team needed ten clickable prototypes of a delivery platform for a large infrastructure programme, to test in workshops with engineers who know the field. The designers built them as plain web pages in one GitHub repository, and the AI coding agent Claude Code did most of the typing.
- The prototypes are deliberately grey and plain, so that experts argue about whether an element should exist and not about its polish.
- Each prototype has a README file recording what was decided, by whom and why. When the AI agent proposed adding an ID line above every record title, the file showed that the label had been removed on purpose.
- Every change to the shared interface kit is logged as pending until someone approves it.
- Anglesea builds each change as a preview copy beside the real page and checks that no content was lost. One tidy-up had dropped a label and a chart key without anyone noticing.
A team that prototypes with an AI agent can keep each decision, and who made it, in a file the agent reads before it changes anything. This is an inference and is not the author's claim; the account covers one project and gives no measured results.
Keep a file beside each prototype that records what was decided, by whom and why, and read it before changing the prototype.
Write a prototype's fidelity rules, such as no new colours or components, in a file that the AI agent reads before each task.
Derived by Working Surface from the article; more in the Playbook. Source line: The code turned out to be the easy part.
8 October 2026: as AI agents produce code and prototypes faster than people can check them, teams must decide where human checks and decisions sit instead of letting review quietly lapse.